YeeYee Beta: an AI editor engineered to spend fewer tokens, not more

Most AI coding tools measure themselves by what the model can do. We spent the beta measuring what it shouldn't have to do — every token the agent burns on something the UI could've just done directly is a token wasted. That discipline is the actual product.

The under-the-hood stuff

Every agent run is capped at 10 steps, on purpose

The loop doesn't wander. It gets 10 iterations to read, edit, and verify — then it stops and hands control back, rather than spiraling through redundant re-reads and "let me also check…" tangents. A single orchestrated step (read → edit → done) gets an even tighter cap. Slower isn't the failure mode we're avoiding — runaway spend is.

A real token budget gate, not a vibes-based one

At 150k tokens spent on a single request, you get a heads-up. At 300k, the loop physically pauses and waits for you to say "keep going." No silent meter climbing in the background — the system stops itself and asks.

The preview button starts the app. The AI doesn't.

This is the one that actually matters: when you click preview, the editor spins up the real dev server directly — no LLM call involved. Every system prompt in the agent explicitly forbids the model from running npm run dev, vite, astro dev, or any serve command to "check its work," because the dev server is already running and hot-reloading. Asking the AI to start your app is asking it to spend tokens duplicating something a button already did for free. We removed that whole category of waste.

Cache-aware by construction

The conversation window is managed to keep a stable prefix instead of shifting on every turn — because a byte-for-byte stable prefix is what makes prompt caching actually work. Get this wrong and you silently re-pay for the same context on every single call. We measured this directly and tuned the eviction policy around it, not around convenience.

Why this is the beta, not a footnote

Anyone can wire an agent loop to a model and call it a coding assistant. The interesting engineering problem is making that loop cheap to run for hours — bounded steps, a real spend gate, and infrastructure (preview, hot-reload) that does its job without asking the model to do it for you. That's the ecosystem: not just an assistant that codes, but one that knows when not to think.

Try it

YeeYee is a local-first Electron app. Beta access details to follow.

Made by dev3lop.com.